Synheart/synheart-emotion-binary-classification-onnx-models

Synheart Binary Emotion Classification Models (ONNX)

ONNX models for on-device binary emotion classification from HRV features.

This repository contains pre-trained ONNX models for inferring momentary affective states (Baseline vs Stress) from Heart Rate Variability (HRV) features. These models are part of the Synheart Emotion SDK ecosystem, enabling privacy-first, real-time emotion inference directly on device.

Made by Synheart AI - Technology with a heartbeat.

Model Performance

ExtraTrees 120_60 Model (trained on WESAD dataset):

• Accuracy: 78.44%
• F1-Score: 72.60%

These models are research-based and trained on the WESAD dataset with Leave-One-Subject-Out (LOSO) cross-validation. The models output probabilistic class scores representing state tendencies, not ground-truth emotional labels.

Confusion Matrices

Performance evaluation results (LOSO cross-validation) for each model configuration:

w120s60_binary (Window: 120s, Step: 60s) - Default Model

w120s5_binary (Window: 120s, Step: 5s)

w60s5_binary (Window: 60s, Step: 5s)

Installation

pip install onnxruntime numpy

Quick Start

import numpy as np
import onnxruntime as ort
import json
from pathlib import Path

class WesadInferenceEngine:
    def __init__(self, model_path, metadata_path=None):
        # Initialize ONNX Runtime session
        self.session = ort.InferenceSession(str(model_path))
        self.input_name = self.session.get_inputs()[0].name

        # Default labels
        self.label_order = ["Baseline", "Stress"]

        # Load metadata if provided
        if metadata_path and Path(metadata_path).exists():
            with open(metadata_path, 'r') as f:
                self.metadata = json.load(f)
                print(f"Model Loaded: {self.metadata.get('model_id', 'Unknown')}")

    def predict(self, hrv_features):
        """
        Predict emotion class from HRV features.
        
        Args:
            hrv_features: list or np.array of 14 features in order:
                ['RMSSD', 'Mean_RR', 'HRV_SDNN', 'pNN50', 'HRV_HF', 'HRV_LF',
                 'HRV_HF_nu', 'HRV_LF_nu', 'HRV_LFHF', 'HRV_TP', 'HRV_SD1SD2',
                 'HRV_Sampen', 'HRV_DFA_alpha1', 'HR']
        
        Returns:
            tuple: (predicted_label, confidence_score)
        """
        input_data = np.array(hrv_features, dtype=np.float32).reshape(1, -1)
        outputs = self.session.run(None, {self.input_name: input_data})
        
        predicted_idx = int(outputs[0][0])
        probabilities = outputs[1][0]
        
        label_name = self.label_order[predicted_idx]
        confidence = probabilities[predicted_idx]
        
        return label_name, confidence

# Usage Example
if __name__ == "__main__":
    MODEL_FILE = "w120s60_binary/models/ExtraTrees.onnx"
    METADATA_FILE = "w120s60_binary/models/ExtraTrees_metadata.json"
    
    # Example HRV features (14 features in correct order)
    example_input = [35.5, 950.2, 55.1, 15.2, 430.0, 620.0, 
                     0.42, 0.78, 1.45, 1450.0, 0.72, 1.3, 1.1, 91.0]
    
    engine = WesadInferenceEngine(MODEL_FILE, METADATA_FILE)
    label, confidence = engine.predict(example_input)
    
    print(f"Predicted Label: {label}")
    print(f"Confidence: {confidence:.2%}")

Repository Structure

├── w120s60_binary/          # Window: 120s, Step: 60s
│   ├── models/
│   │   ├── ExtraTrees.onnx
│   │   └── ExtraTrees_metadata.json
│   └── figures/
│       └── confusion_matrix_ExtraTrees_loso_binary.png
├── w120s5_binary/           # Window: 120s, Step: 5s
│   ├── models/
│   │   ├── ExtraTrees.onnx
│   │   ├── LogReg.onnx
│   │   ├── RF.onnx
│   │   └── [metadata files]
│   └── figures/
│       ├── confusion_matrix_ExtraTrees_loso_binary.png
│       ├── confusion_matrix_LinearSVM_loso_binary.png
│       ├── confusion_matrix_LogReg_loso_binary.png
│       ├── confusion_matrix_RF_loso_binary.png
│       └── confusion_matrix_XGB_loso_binary.png
└── w60s5_binary/            # Window: 60s, Step: 5s
    ├── models/
    │   ├── ExtraTrees.onnx
    │   └── ExtraTrees_metadata.json
    └── figures/
        └── confusion_matrix_ExtraTrees_loso_binary.png

Input Features

The models expect 14 HRV features in this exact order:

1. RMSSD
2. Mean_RR
3. HRV_SDNN
4. pNN50
5. HRV_HF
6. HRV_LF
7. HRV_HF_nu
8. HRV_LF_nu
9. HRV_LFHF
10. HRV_TP
11. HRV_SD1SD2
12. HRV_Sampen
13. HRV_DFA_alpha1
14. HR

Output

• Labels: "Baseline" or "Stress"
• Confidence: Probability score for the predicted class

About Synheart

Synheart is an AI company specializing in on-device human state intelligence from physiological signals. Our emotion inference models are part of a comprehensive SDK ecosystem that includes:

• Multi-Platform SDKs: Dart/Flutter, Python, Kotlin, Swift
• On-Device Processing: All inference happens locally for privacy
• Real-Time Performance: Low-latency emotion detection from HR/RR signals
• Research-Based: Models trained on validated datasets (WESAD)

Using with Synheart Emotion SDKs

For production applications, we recommend using the official Synheart Emotion SDKs which provide:

• ✅ Complete feature extraction pipeline (HR/RR → 14 HRV features)
• ✅ Thread-safe data ingestion
• ✅ Sliding window management
• ✅ Unified API across all platforms
• ✅ Integration with Synheart Core (HSI)

SDK Repositories:

• Python SDK - pip install synheart-emotion
• Dart/Flutter SDK
• Kotlin SDK
• Swift SDK

Direct ONNX Usage

This repository is ideal if you:

• Want to use the models directly with ONNX Runtime
• Have your own feature extraction pipeline
• Need to integrate into custom inference systems
• Are doing research or experimentation

Links

• Synheart AI: synheart.ai
• Synheart Emotion: Main Repository
• Documentation: Full SDK Documentation

License

These models are part of the Synheart Emotion ecosystem. Please refer to the main Synheart Emotion repository for licensing information.

---

Made with ❤️ by Synheart AI